How social change brings bias to predictive models used in criminal justice settings

Robert J. Sampson, the Woodford L. and Ann A. Flowers University Professor, co-authored a new study on risk assessment instruments (or RAIs). Kris Snibbe/Harvard Staff Photographer

3 min read

New research has uncovered a surprising bias in the risk assessment instruments (or RAIs) widely used in criminal justice settings. It turns out, their predictive powers are quick to erode. Social scientists found that RAIs trained on one birth cohort vastly overestimated the probability of arrest in late adolescence and early adulthood for a slightly younger birth cohort.

“This is not just a minor quirk,” stressed Robert J. Sampson, the Woodford L. and Ann A. Flowers University Professor. “We’re overpredicting their probability of arrest by up to nearly 90 percent.”

“Overprediction also occurs for all racial groups,” Sampson added. “What that means is this is a unique form of bias. It’s what we call in the paper cohort bias.”

The study was published this week in the Proceedings of the National Academy of Sciences. Sampson and his co-authors specifically scrutinized RAIs trained on individuals born between the years 1979 and 1988 to predict the likelihood of arrest between the ages of 17 and 24. The same algorithmic tools were then applied to those born in the mid-1990s. This approach was made possible by the Project on Human Development in Chicago Neighborhoods, which Sampson helped launch in the mid-1990s to follow various birth cohorts from youth to adulthood.

“If you’re predicting if someone will be arrested — or whether they should be given bail, get treatment, or go to prison — these are all based on assessments of the future,” Sampson explained. “But they are reliant on observed behavior from the past.”

Those born in the 1980s grew up with high crime rates and the war on drugs, while those born in the ’90s came of age amid rapidly falling crime rates and changing policing patterns. Individual risk factors — like growing up in poverty or in a single-parent household — also shifted in how they predicted future arrest. As a result of these and other societal changes, a range of prediction models trained on the study’s older set substantially overpredicted the probability of arrest for the younger cohort.

This suggests that outdated RAIs can perpetuate unfair and unnecessary encounters with the criminal justice system.

The research was not meant to dismiss the role of prediction, Sampson emphasized. “Our aim is to understand when it goes wrong and how we might counteract that. And the takeaway is that social change really degrades the accuracy of prediction.”

Also outlined are possible solutions, including more frequent updates to information systems, which are surprisingly old in many jurisdictions. As an example, the paper cites one used by the New York City Criminal Justice Agency to predict pretrial risk of nonappearance. It was developed in 2003 and used without updating until 2020.

The research described in this report was supported by the Office of Juvenile Justice and Delinquency Prevention.